Avaliação do uso de uma rede neural artificial e regressão linear múltipla para previsão do teor de silício no ferro-gusa utilizando variáveis operacionais do alto-forno.

Abstract

The increasing competition in the steel sector has led steel producers to look for different ways to improve the quality of their products. The delivery of hot metal with a chemical composition as expected for the steel shop is of fundamental importance for this objective. In this context, one of the main chemical elements used as an indicator of the quality of hot metal is silicon. Therefore, many studies have been developed over the decades focusing on the predictability and control of silicon content in hot metal. Recently, the application of computational models for this task is increasing, especially the use of artificial neural networks. In the present work, the use of a Multilayer Perceptron neural network (MLP) was evaluated in the construction of a model that used operational variables as input data and compared it with another built from a multiple linear regression. To obtain the results of the multiple linear regression and the neural network, Excel and Matlab software were used, respectively. For the first, all data was used as input and for the second, the data was divided into training and test sets, being randomly selected 100 times for each hyperparameter configuration experimented. The artificial neural network obtained an average correlation coefficient (R2) of 0.2615, which is higher than that obtained by multiple linear regression of 0.0971, demonstrating that it has a better ability to predict the silicon content in hot metal in front of the variables presented.